Machine learning-based prediction of low-value care for hospitalized patients

Area of Science:

• Healthcare Informatics
• Clinical Decision Support
• Machine Learning in Medicine

Background:

• Low-value care, defined as costly treatments with minimal benefit, is a significant issue in US hospitals.
• Traditional methods to reduce low-value care have limited success.
• Behavioral science principles can inform novel strategies for care improvement.

Purpose of the Study:

• To develop and evaluate machine learning models for predicting low-value care at the point of care.
• To create a targeted prompting system to suggest alternative treatments before clinical decisions are finalized.
• To assess the feasibility of using machine learning to reduce the administration of intravenous albumin for fluid resuscitation in intensive care unit (ICU) patients.

Main Methods:

• Utilized electronic health record data from a multi-hospital system to identify low-value care practices, specifically albumin administration in ICUs.
• Developed two machine learning models: a single-stage model predicting direct albumin administration and a two-stage model predicting fluid administration followed by albumin.
• Defined clinical features at 4-hour intervals within ICU episodes for model training and validation.

Main Results:

• Examined 87,489 ICU episodes (approx. 1.5 million 4-hour periods).
• Both single-stage and two-stage models achieved an area under the receiver operating characteristic curve of 0.86.
• The models demonstrated a positive predictive value of 0.21-0.22, with the potential to prevent 10% of albumin administrations and prompt physicians every 4.2 days.

Conclusions:

• Predicting low-value care using machine learning is feasible and can support point-of-care prompting systems.
• Targeted prompting systems can offer timely suggestions to clinicians before decisions are made.
• A two-stage prediction model offers calibration flexibility without significantly improving predictive performance over a single-stage model.